BRENNEN

Industries / Technology / Electronics and semiconductors

Yield and inspection

Yield and inspection is the work of finding defects on masks and wafers, measuring what was made, and using that data to raise the share of chips on each wafer that work, a share called the yield.

6

Milestones achieved, from 1964 to 2019.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1964

    Yield gets a formula

    B. T. Murphy of Bell Labs linked a chip's area to its yield with a model that lets the number of defects vary from chip to chip, and used it to find the chip size with the lowest cost per component.

    Lessonlarger chips catch more defects, so chip size is an economic choice as well as a technical one.

    manuscript.isc.ac

    Credit

    B. T. Murphy, Bell Telephone Laboratories

  2. 1975

    Machines inspect photomasks

    One flaw on a photomask prints onto every chip made with it. Ken Levy began raising money for KLA in 1975, and its first product was a machine that "optically viewed the photomask and determined if it was free of defects".

    archive.computerhistory.org

    Credit

    Ken Levy and Bob Anderson, KLA

  3. 1983

    Yield theory becomes a tool

    IBM engineers derived yield statistics from how defects and faults form during manufacturing, which pulled earlier formulas into one practical model for controlling production. Later IBM work built in the "long observed" clustering of defects, which bunch together instead of spreading evenly.

    manuscript.isc.acsearch.isc.ac

    Credit

    Charles H. Stapper, Frederick M. Armstrong and Kiyotaka Saji, IBM

  4. 1998

    Factories copy, not reinvent

    Under Intel's Copy Exactly method, a new factory matches its development factory in equipment, settings and process; a 1998 Intel paper said that unchecked engineering changes slow output when a process moves into manufacturing. Craig Barrett traces the method to Intel's mid-1980s manufacturing turnaround.

    Lessonfreezing a recipe that works can raise output more than local improvements do.

    search.isc.acchiphistory.org

    Credit

    Intel (R. A. Gasser Jr.; Craig Barrett)

  5. 2015

    Wafer maps go public

    A wafer map shows which chips on a wafer passed or failed testing, and its patterns point to what went wrong. Researchers in Taiwan released WM-811K, a set of 811,457 real wafer maps, which gave people outside chip factories production data to test pattern-recognition methods on.

    manuscript.isc.ac

    Credit

    Ming-Ju Wu, Jyh-Shing Roger Jang and Jui-Long Chen

  6. 2019

    Bad resist scraps wafers

    TSMC found that a batch of photoresist from a supplier contained an abnormally treated ingredient that damaged 12 nm and 16 nm wafers at its Fab 14B, and expected the incident to cut its first-quarter revenue by about $550 million. TSMC responded by strengthening inspection of wafers during processing and tightening control of incoming materials.

    Lessonchecking incoming materials is part of yield work, not only checking wafers.

    pr.tsmc.com

    Credit

    TSMC (disclosure)

Watch for

  1. AI defect sorting at TSMC

    TSMC and NVIDIA say TSMC uses NVIDIA's Metropolis vision-AI software and TAO Toolkit to classify wafer defects, finding nanometre-scale defects better and needing less retraining when tools or defect types change. No figures were given.

    Achieved whenTSMC or an independent party publishes measured production results, such as classification accuracy or a change in yield or review time.

    itbrief.news

    Claimed by

    TSMC and NVIDIA, June 2026

  2. Electron-beam review for packaging

    Applied Materials says its SEMVision G7AP uses an electron beam to review and sort defects on the silicon, organic and glass bases used in advanced chip packaging, and that it is already in production at unnamed memory and logic makers.

    Achieved whena named chipmaker or an independent party confirms production use and reports results.

    siliconsemiconductor.net

    Claimed by

    Applied Materials, June 2026

Open

  1. 01

    Shuttle chip yield record

    A public record of electrical test results from at least 50 chips from one Tiny Tapeout run, giving for each chip whether the run's factory-test design passes and the frequency of a ring oscillator, a loop of inverters whose speed shows how fast that chip's transistors came out. Add an analysis of the pass rate and of the spread in speed from chip to chip. (The design-by-design test under Packaging and test can share boards and data.)

    The brief
    Credit · unclaimed
  2. 02

    Mask-defect printability set

    A public, simulation-labelled dataset of photomask defects: layout clips with added defects of known type, size and place (pinholes, extensions, bridges, missing corners), each simulated across a range of focus settings and labelled "printable" or "not printable" by a rule stated in advance, such as a change in printed line width above 10%. Then train a model that predicts printability from the mask image alone, a call factories now make with aerial-image review tools such as Zeiss AIMS.

    The brief
    Credit · unclaimed

Open 01 · the brief

Shuttle chip yield record

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public record of electrical test results from at least 50 chips from one Tiny Tapeout run, giving for each chip whether the run's factory-test design passes and the frequency of a ring oscillator, a loop of inverters whose speed shows how fast that chip's transistors came out. Add an analysis of the pass rate and of the spread in speed from chip to chip. (The design-by-design test under Packaging and test can share boards and data.)

Why it is still open

Tiny Tapeout's own posts say only that the team "confirmed our test designs work", and the wafer.space pages that track these runs give figures "measured from the public layout", not from chips. Per-chip results show up only as scattered comments, such as two users reporting about 510 MHz and 517 MHz for one ring oscillator.

tinytapeout.comwafer.spacetinytapeout.com

Built with what exists

Tiny Tapeout demo boards with the MicroPython SDK and its microcotb test tool; the run's factory-test design; ring-oscillator projects on the same run; Python and pandas; a shared form or Git repository for chip owners to submit results.

  1. First step

    Write a MicroPython script that runs the factory test and reads one ring oscillator, run it on one board, and post it with a submission form in the Tiny Tapeout community.

  2. Done looks like

    results from at least 50 chips from one named run, each with board ID, date, supply voltage, factory-test pass or fail and ring-oscillator frequency, published as a CSV with a notebook that reports the pass rate with a 95% confidence interval and the mean and spread of frequency; a second person reruns the notebook and gets the same numbers.

Open 02 · the brief

Mask-defect printability set

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public, simulation-labelled dataset of photomask defects: layout clips with added defects of known type, size and place (pinholes, extensions, bridges, missing corners), each simulated across a range of focus settings and labelled "printable" or "not printable" by a rule stated in advance, such as a change in printed line width above 10%. Then train a model that predicts printability from the mask image alone, a call factories now make with aerial-image review tools such as Zeiss AIMS.

Why it is still open

A December 2025 paper states that "Lithography defect data from semiconductor industry are rarely accessible for research use", and the set it built covers wafer defects, not mask defects. We found no public mask-defect printability dataset.

arxiv.org

Built with what exists

layout clips from LithoBench; OpenILT's lithography simulator (PyTorch); EUVlitho for an optional EUV subset; KLayout or shapely to draw defects; PyTorch for the classifier.

  1. First step

    Add one pinhole at five sizes to 100 LithoBench clips, simulate them, and plot the change in printed width against defect size to fix the labelling rule.

  2. Done looks like

    at least 20,000 defect cases across at least four defect types and a sweep of sizes, each with its geometry, simulated prints and label; a published classifier that reaches stated recall and false-alarm rates, such as 95% recall on printable defects with at most 10% false alarms, on layouts it never saw in training; everything rebuildable from scripts.